Abstract:Based on false nearest neighbors and unscented kalman filter (FNN-UKF), a dynamic prediction method for alumina concentration was proposed. In the new KPCA feature subspace, it was inspired by FNN that interpretation of alumina concentration would be estimated by calculating the variables mapping distance in the KPCA space to select secondary variables. Selected variables were introduced into BP neural networks as input vector. UKF algorithm, in which estimated value and variance matrix of state were updated to improve the generalization capability of the networks, was used to train weight values and threshold values. By using 247 samples of 160KA operating aluminum cell from a factory, experimental results demonstrated that the forecast error of 228 samples was ±1%, the computation was decreased to 52.07%.The method in which the computation time was reduced effectively can surely accuracy of parameter estimation.